Content
92%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
A dense, operational audit playbook: concrete grep commands and per-agent path tables, a clearly sequenced Step 0–6 workflow with a findings validator and explicit triage and closure rules, and disciplined offloading of full detail to real, well-signaled reference files. The only deduction is minor token overhead in the motivating intro, the cross-reference paragraph, and the partially duplicated headline location table.
Suggestions
Trim the opening statistics ('of 31,132 public skills analysed, 26.1% had a vulnerability and 5.2% looked malicious') to a single clause or move them to a reference file — the model does not need motivating data to execute the audit.
Compress the Step 1 headline location table, which overlaps references/locations.md, to agent names plus a pointer, or keep only the columns (source, pin, runtime) not already tabulated there.
Shorten the cross-references paragraph (ten sibling skills) to the three most likely handoffs (prompt-injection, secrets-audit, incident-triage) and move the full map to a reference file.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is information-dense and assumes Claude's competence (it never explains what MCP servers or hooks are), but a few passages could be trimmed: the motivating statistics ('of 31,132 public skills analysed, 26.1% had a vulnerability and 5.2% looked malicious'), the long sibling-skill cross-reference paragraph, and the Step 1 headline location table that partially duplicates references/locations.md. This matches 4 ('efficient; minor instances of over-explanation that could be trimmed') rather than 5, whose every token must earn its place; it is clearly above 3, which requires genuinely unnecessary explanation or padding. | 4 / 5 |
Actionability | Fully executable guidance throughout: the Step 2 grep table gives concrete `rg` patterns (with 'Set `T` to the inventoried paths first' and a note on unescaping `\|`), Steps 3–4 give copy-paste bash blocks over named settings files, and the skill specifies the exact SkillSpector invocation (`skillspector scan --no-llm .claude/skills -f json -o skillspector.json`), the pinned MCP Inspector snapshot command, the full findings.json record fields, and the validator command `node references/validate-findings.mjs findings.json`. Matches the 5 anchor; 4 would tolerate missing key details, and none are missing. | 5 / 5 |
Workflow Clarity | Steps 0–6 are clearly sequenced (authorization/scope → inventory → content checks → hooks and settings → MCP servers → optional scan → verdicts and output) with explicit validation checkpoints: 'Fix every error before writing prose' after running the findings validator, explicit hit-triage rules ('Directive vs documentation,' 'Concealment is the signal'), and a confirmed / needs_validation / rejected closure model with per-item validation plans. Matches the 5 anchor (explicit validation, feedback loops, checklists); the missing-validation cap at 3 does not apply because validation is present and the skill is read-only. | 5 / 5 |
Progressive Disclosure | Each section keeps only headline detail inline and clearly signals its one-level-deep reference — 'The per-agent path tables and a one-shot discovery command are in references/locations.md,' 'the full set, with portable grep / perl equivalents, is in references/content-checks.md,' plus browser-runtimes.md, launch-directories.md, ide-extensions.md, findings-format.md, and severity-calibration.md — and all referenced files exist in references/. The two bundle files not named in the body (findings.schema.json, to-sarif.mjs) are documented in findings-format.md, so nothing is orphaned. Matches the 5 anchor (clear overview, well-signaled one-level-deep references, appropriate split); 4 would show organization gaps, and none are evident. | 5 / 5 |
Total | 19 / 20 Passed |